Abstract: Knowledge Graph Completion aims to solve the incompleteness problem of the knowledge graph by inferring the missing triples based on the triples already known in the knowledge graph. Most of the existing methods directly fuse entities and relations to extract features, ignoring the common features and specific attributes between the triples. To this end, the paper proposes ConvPair, a convolutional neural network model based on pairwise relation vectors, to improve the link prediction performance of Knowledge Graph completion through pairwise relation modelling, feature fusion optimization, and relation-aware convolutional operations. The key idea of the ConvPair model is to split the relation r into a head relation rh and a tail relation rt, and rh is fused with the head entity h, and then a commonality feature encoding module to obtain the common feature entity f. Next, f is fused with rt to predict the tail entity t through the specific attribute extraction module. In terms of entity-relationship fusion, a checkerboard reshaping operation is performed on the vector representations of h and rh to fully fuse the features; at the same time, the vector representations of f and rt are directly spliced together, preserving the ternary translation property in the specific attribute extraction. Finally, a relation-aware convolutional kernel is designed to effectively handle complex relations and extract relation-specific features. Experimental results show that ConvPair outperforms existing state-of-the-art models on datasets such as FB15k-237, WN18RR and YAGO3-10, demonstrating excellent performance and generalizability.

Key words: knowledge graph completion, paired relation vectors, convolutional neural network, link prediction

摘要： 知识图谱补全旨在依据知识图谱中已知三元组推理出缺失的三元组，以解决知识图谱的不完整性问题。现有方法大多将实体和关系直接融合提取特征，忽略了三元组中之间的共性特征和具体属性。为此，该文提出了一种基于成对关系向量的卷积神经网络模型ConvPair，通过成对关系建模、特征融合优化和关系感知卷积操作，提升知识图谱补全的链接预测性能。ConvPair模型的核心思想是将关系r拆分为头关系rh和尾关系rt，rh与头实体h融合后，通过共性特征编码模块得到共性特征实体f。接着，f与rt进行融合，通过具体属性提取模块来预测尾实体t。在实体与关系的融合方面，对h和rh的向量表示进行棋盘重塑操作，以充分融合特征；同时，直接拼接f和rt的向量表示，保留在提取具体属性时的三元组翻译特性。最后，设计关系感知卷积核，有效处理了复杂关系并提取关系的特定特征。实验结果表明，ConvPair在FB15k-237、WN18RR、YAGO3-10等数据集上优于现有先进模型，展现了卓越的性能和泛化能力。

关键词: 知识图谱补全, 成对关系向量, 卷积神经网络, 链接预测